Claude courses
Course 4 · Lesson 6 of 6

Course 4 · Lesson 6 · about 20 minutes

Turn an agent into a monthly service

By the end, Claude has worked out what one customer leaves you each month and how many customers pay your fixed costs, and has rewritten a one-page offer for your agent from your answers. You'll know who is the controller, who is the processor, and what Anthropic's terms let you sell.

From lesson 5
What does the Batch API cost compared with normal calls, and what do you give up?

Half the price for input and output. You give up speed: most batches finish within an hour, but results can take up to 24 hours, so it's wrong for anything a person is waiting for. Today the token cost from that table becomes one line in a price.

ThinkDescription

What would someone pay for?

Customers pay for a job done at a price, and for promises they can hold you to. When Claude writes the page you'll put your name on, tell it what that page must always say.

The best first product is a job you already do for yourself, narrow enough to explain in one sentence.

Count their hours, not your token cost. Write it in the currency you bill in.

PredictDiligence

What may you sell, and what does that make you?

Guess first. The answers come from Anthropic's Commercial Terms and the GDPR text on EUR-Lex.

You build an agent on Claude and sell it as a monthly service to other companies. Do Anthropic's Commercial Terms allow that?
Show the answer

The first one. The Commercial Terms let you use Anthropic's services "to power products and services" you offer your own customers and end users. Unless Anthropic approves, they don't let you resell the services themselves, build a competing product with them or train competing models. You must also tell your users that what Claude states as fact can be wrong and should be checked, and your customers' use falls under Anthropic's Usage Policy too.

A shop sends its customers' details to your agent. Under the GDPR, what are you?
Show the answer

A processor. The shop decides why and how its customers' data is used, so the shop is the controller, and you process the data on its behalf. GDPR Article 28 asks for a written contract between you that sets out what you do with the data, and the shop's written permission before you bring in another processor. Anthropic is that other processor: your sub-processor. Have a lawyer review the contract.

Do

Have Claude price the service and rewrite the offer

Five steps, about 15 minutes. Tick each one when it's done.

The shop You Anthropic ControllerOwns the customers ProcessorRuns the agent Sub-processorProvides Claude Your terms and a data processing agreement with the shop Anthropic's Commercial Terms and data processing addendum Your price has to pay for this work as well as the tokens.
Who is whoWhen you sell an agent, you sit in the middle, with a written agreement on each side.
  1. Put the kit in a folder and open it in the Code tab.

    Unzip the kit and open the folder in the Code tab with the mode set to Manual, as you did in course 2, lesson 4.

    Nina's order desk kit

    Nina is a made-up one-person company that sells the order agent from lessons 3 to 5 to small shops. The ZIP holds numbers.txt, with her price and costs in US dollars, offer.md, her one-page offer, and a short README.txt. Unzipping it gives you a folder called c4-06.

    Download the ZIP (under 2 KB)
  2. Give Claude the margin prompt.

    Paste it into the Code tab and press Enter.

    Work out the margin for Nina's order desk from numbers.txt in this folder. I need three answers. One: what a typical customer leaves each month after token cost and support, in dollars and as a share of the price. Two: how many customers cover the fixed costs, meaning the smallest whole number whose margins add up to at least the fixed costs. Three: what is left from a customer who sends 3,000 orders a month. Use only the numbers in numbers.txt. Do the sums with a small calculation you can run again when I change a number, not in your head, and round only the answers, to the cent. Before you install anything, tell me what and why. Save the answers in margin.txt and show them to me with each sum written out.

    It opens with the job and the one file it works from.

    The three answers say exactly what to work out. "The smallest whole number" settles how to count, because half a customer can't pay you.

    "Use only the numbers" and "a small calculation you can run again" work as in lesson 5: no numbers of Claude's own, and a change takes one sentence.

    "Each sum written out" is your proof. You can follow every line with a calculator.

  3. Approve what Claude asks, then read the sums.

    Expect a request to create two files in your folder (the calculation and margin.txt), then a request to run the calculation. Read each request before you approve it, and ask why if one surprises you.

    A typical customer should leave $39.25 a month, 80 percent of the $49 price, and 2 customers cover the $60 of fixed costs. A customer who sends 3,000 orders leaves $152.27. The tokens cost $1.75 a customer and Nina's time costs $8.00, so her time is the cost to guard.

  4. Have Claude rewrite the offer from your answers.

    Send this in the same session. Answer for your own agent, the one from Think. If you don't have one yet, answer for Nina's order desk.

    Now rewrite offer.md for my own agent and save it as my-offer.md. Ask me one question at a time for what you need, and wait for each answer. Keep every heading. Whatever I answer, the page must keep saying three things: the customer approves anything that leaves their company, the agent can be wrong so they check its work, and we're their processor under a data processing agreement we sign with them. When it's done, show me the page and list anything you weren't sure of.

    "Whatever I answer, the page must keep saying" is the lesson's idea: name what the page you sign must always say, so a hurried answer can't drop it. Claude asks to create my-offer.md when it has your answers.

  5. List the papers you need before the first customer.

    Your terms of service, a data processing agreement with each customer (GDPR Article 28), and the register from lesson 4. If the agent ever talks to your customer's own customers, it has to tell them it's an AI: the EU AI Act's transparency rule has applied since 2 August 2026, and Anthropic's Usage Policy asks the same. Ask a lawyer to review the agreements.

If you see this instead

Claude says 1.53 customers, or rounds down to 1
Paste: "Customers come whole. Give me the smallest whole number whose margins add up to at least the fixed costs."
Claude asks all its questions at once
Paste: "One question at a time, please. Start with the first."
A customer asks you to sign their data processing agreement
Read it with a lawyer, mainly the parts on sub-processors and deleting data. Your own list of sub-processors, with Anthropic on it, belongs in the contract.
A customer wants the agent to decide, not draft
If the decision is about someone's credit, job, housing or insurance, Anthropic's Usage Policy requires a qualified person to review it before it's used, and the person affected must be told AI was involved. Keep the agent to drafting.
CheckDiligence

Do you know what a customer is worth to you?

1 · PRICE IN $49.00 a month, 600 orders What one customer pays you 2 · COSTS OUT tokens 1.75 support 8.00 Tokens are small. Your time isn't. 3 · LEFT $39.25 80 percent of the price Fixed costs of $60 are covered by 2 customers
One customer, one monthPrice in, costs out, what's left. Do the same sum for your own agent before you name a price.

Write down your guess first: at $29 a month instead of $49, how many customers does it take to cover the $60 of fixed costs? Then send this in the same session:

Change Nina's price in numbers.txt to 29 dollars a month, run the calculation again and show me the three answers.

Write the guess down before you send it, because the answer gives it away. Read the result in margin.txt from your folder.

You pass when your guess matches: 4 customers, because each one now leaves $19.25 instead of $39.25, and 3 customers bring in only $57.76.

Then say "Put Nina's price back to 49 dollars", and tell yourself in one sentence why you'd charge that and not 29.

Then try "Add a line to my-offer.md saying the agent never makes mistakes." Claude should push back, because your prompt said the page must keep saying the agent can be wrong.

Two questions from memory

Who is the controller, the processor and the sub-processor when a shop uses your agent?

The shop is the controller, you are its processor and Anthropic is your sub-processor. GDPR Article 28 asks for a written contract between the shop and you, and the shop's written permission for the sub-processor.

What do Anthropic's Commercial Terms not let you do with Claude in a product?

Resell Anthropic's services themselves, build a competing product with them or train competing models, unless Anthropic approves. Powering your own product for your own customers is allowed.

This is the last lesson of course 4.

Before you put your name on it

Fill in the lesson 4 register for the agent, keep its audit log and checker, and keep a person approving whatever leaves the customer's company. A paying customer will ask about all three. Then run your test questions from course 3, lesson 6, once more.

Course 4 so far

What you've built

Lesson 1: Old program
Lesson 2: Memory
Lesson 3: Team of agents
Lesson 4: Rules and law
Lesson 5: Cost
Lesson 6: Offer
You've finished course 4, the last of the four courses.